Comparative Analysis of Sensor Performance in Autonomous Vehicles: Evaluating RGB, Thermal Cameras, and LiDAR Under Adverse Weather Conditions

Reliable perception under adverse weather conditions remains a critical challenge for autonomous vehicles. This paper presents a controlled comparative study of three key sensing modalities: RGB cameras, thermal cameras, and LiDAR, for object detection across diverse weather and illumination conditions. Using the Seeing Through the Fog dataset, we construct train/test splits, provide new manual annotations for thermal test data, and employ state-of-the-art detectors, including YOLOv10x for image modalities and CasA+ for LiDAR point clouds.We evaluate both specialist models trained for each weather condition and a general model trained on data from all conditions. Performance is analyzed quantitatively using mean average precision at multiple intersection over union thresholds, complemented by LiDAR-specific quality metrics, and qualitatively through detailed visual comparisons. Results demonstrate that RGB imaging achieves the highest accuracy in clear conditions but suffers significantly at night and in fog. Thermal imaging excels in low-light detection of pedestrians and objects with active heat sources, but struggles with boundary precision, class confusion, dealing with cold objects, and scattering from snow or rain. LiDAR provides stable geometry and depth cues under varied illumination; however, performance degrades in fog and rain, and at long range due to sparsity. Collectively, the analysis highlights that no single sensor is sufficient for all conditions, where each performs better in different scenarios. Instead, the complementary strengths of the three modalities motivate the use of multimodal fusion strategies to ensure robust perception in safety-critical environments.